Causal Inference

This course explores two complementary formal approaches to causal reasoning. We begin with an introduction to causal graphical models, examining how Causal Bayes Nets and Structural Causal Models capture causal structure and support a variety of inferences, with a focus on the formal distinction between ‘seeing’, ‘doing’ and ‘imagining’. In subsequent classes, we make our way up Pearl’s ladder of causation, first discussing causal learning from observational and interventional data, and then exploring various aspects of counterfactual reasoning: how one can evaluate counterfactual conditionals, and the various problems involved in judging what caused a particular event (‘actual causation’).

The second part of the course shifts to representing causal knowledge with structured programs. We examine probabilistic program induction as a framework for causal reasoning, treating the acquisition of causal knowledge as a search problem over program spaces. Drawing on formal methods including probabilistic context-free grammars (PCFGs), approximate Bayesian inference, and adaptor grammars, we explore how structured representations enable few-shot learning and generalization. The course concludes with recent developments in causal discovery through active learning and applications to generative agents, highlighting open questions at the intersection of causal cognition and artificial intelligence.